{"id":"2c5b7cbb-3094-4485-9401-7a8adf359abe","arxiv_id":"2607.16282","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 22k-parameter, 7.6-MFLOP stateless detector fed ego-motion-normalized residual-motion channels tracks small aerial persons at 31.85 FPS on a Pi Zero 2W, beating YOLOv8n by 2.7x AP25 with 16x lower latency.","lead":"EMTS-Det is a five-stage drone-person tracker that analytically removes the camera's own motion, then hands the residual-motion evidence to a 22k-parameter detector and a verification network, so no learned temporal machinery is needed. On a Raspberry Pi Zero 2W it reports 31.85 FPS with 0.462 AP25 on 1,000 real UAV videos, versus 1.95 FPS and 0.172 AP25 for YOLOv8n.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generalization to real footage rests on synthetic-motion pseudo-clips with hand-set trajectory parameters, and the sole real-video accuracy table lacks a stated labeling protocol; the transfer claim is not yet audited.","rationale":"The reader's weakest assumption identifies exactly the point I find most load-bearing: the headline generalization evidence is measured under an author-defined motion regime. VisDrone-DET is a still-image dataset; converting it to pseudo-clips with hand-set trajectory parameters tests whether the system can exploit synthetic motion built from those parameters, not whether it generalizes to real UAV ego-motion and real person articulation. The in-domain composite bench cannot validate this because it uses the same generative process. The one real-corpus table (Table 10) would resolve the question, but its labeling protocol is unstated, so it cannot currently be audited. I do not think this warrants rejection: the paper is unusually honest, reports multi-seed ablations with variance, documents calibration failures, measures on-device latency, and explicitly lists field-labeled benchmarks as future work. The concern is about evidence strength, not about internal inconsistency or suspected fabrication. A clean test on real video with official labels, or hand labels on a subset of the 1,000-video corpus, would either confirm the transfer claim or expose the pseudo-motion gap. This matches the reader's conditional verdict, so no verdict adjustment is needed.","tokens_in":22685,"tokens_out":3726,"duration_ms":43544,"concrete_test":"Evaluate the deployed int8 pipeline and the identically fine-tuned YOLOv8n baseline on genuine consecutive-frame drone video with official human-annotated boxes, e.g., a held-out subset of VisDrone-VID sequences whose images are disjoint from the 85% training split, using the same AP25 protocol. If EMTS-Det's real-video AP25 falls substantially below the 0.444 pseudo-clip figure, or the gap versus YOLOv8n narrows, the synthetic-motion recipe is the source of the claimed transfer. As a minimal cross-check, hand-label a random sample of 50-100 videos from the Table 10 corpus and recompute AP25; reproducing 0.462 with human labels would settle the labeling-protocol objection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ego-normalized temporal evidence in the input representation transfers to real drone imagery, supported by 0.444 AP25 on 'held-out real VisDrone-DET' versus 0.103 for fine-tuned YOLOv8n (Tables 4-5, Sec. 6.5). But this evaluation converts static VisDrone images into 4-frame pseudo-clips through the deployed Stage A/B code with hand-set trajectory parameters: up to 4 px/frame drift, 0.008 rad/frame roll, 0.6% scale rate, per-person lower-body shear (Sec. 4.0.2). The motion channel therefore encodes the authors' model of drone/person motion, not actual drone motion. The transfer claim assumes this synthetic recipe reproduces the task-relevant statistics of real UAV footage. The composite bench shares the same generative process, so in-domain scores cannot validate this. The only accuracy evaluation on genuine real videos (Table 10, 1,000 UAV videos) reports AP25/AP50/Recall but never states the labeling protocol; if labels are automatic (e.g., YOLO-derived), the comparison could be systematically biased. Limitations 2 and 4 explicitly concede that a field-collected labeled benchmark and hand-labeled distant-regime ground truth remain outstanding. Until the pseudo-motion assumption or the Table 10 labeling is independently verified, the headline '0.444 on real held-out VisDrone' and '0.462 on 1,000 real videos' overstate what is currently demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EMTS-Det, a five-stage aerial person-tracking pipeline for milliwatt-class hardware. Stage A estimates ego-motion as a 4-DoF similarity transform; Stage B forms ego-motion-normalized residual-motion channels; Stage C is a 21,941-parameter, 7.6-MFLOP stateless CenterNet-style detector; Stage D is a Kalman tracker in stabilized coordinates; Stage E is a 1-D CNN verifier of human motion. Training uses a synthetic-motion curriculum with motion channels generated by the deployed Stage A/B code. The central claim is that temporal evidence belongs in the input representation, computed analytically, rather than in learned temporal machinery, and that this representation transfers to real drone imagery. Evidence includes multi-seed ablations on a composite bench and a held-out VisDrone-DET split, int8 quantization-failure analysis, an on-device Raspberry Pi Zero 2W evaluation over 1,000 real UAV videos, and a field sequence.","tokens_in":22946,"tokens_out":4725,"duration_ms":54014,"significance":"If the central claim holds, the paper is significant: it demonstrates a 1,100x compute reduction over YOLOv8n while maintaining or improving small-person detection on real imagery, provides a concrete negative result for learned temporal modules (TSM) once motion channels are available, and documents reproducible int8 calibration failure modes for stateless and stateful detectors. The paper is disciplined in several ways: ablations are repeated over three seeds with standard deviations; YOLOv8n is fine-tuned on the identical training mixture; arithmetic is internally consistent; limitations are admitted; and the authors release scripts for training and on-device measurement. However, the load-bearing generalization evidence currently depends on synthetic motion applied to static VisDrone images, and the only genuine real-video accuracy table lacks a stated labeling protocol. These gaps must be closed or substantially reframed before the transfer claim can be accepted as demonstrated.","major_comments":[{"comment":"The held-out VisDrone-DET generalization evidence is obtained by converting static images into 4-frame pseudo-clips using the authors' Stage A/B code with hand-set trajectory parameters: up to 4 px/frame drift, 0.008 rad/frame roll, 0.6% scale rate, and per-person lower-body shear. The motion channels therefore encode the authors' model of drone/person motion, not measured drone motion. The headline 0.444-vs-0.103 comparison and the 6x collapse of the luminance-only variant in Table 5 are conditional on this synthetic recipe reproducing the task-relevant statistics of real UAV footage. Limitation 2 explicitly concedes that a field-collected labeled benchmark is still needed. Please add a sensitivity analysis over the trajectory parameters, or better, evaluate on real videos with hand labels, before presenting 0.444 as evidence of transfer to real imagery.","section":"Sec. 6.5, Sec. 4.0.2, Tables 4-5"},{"comment":"The 1,000-video real-world evaluation on Raspberry Pi Zero 2W is the only accuracy result on genuine videos, but the labeling protocol is not stated. Who created the ground-truth boxes, how were tiny/distant persons annotated, were labels manual or machine-generated, and what quality control was used? Without this information the AP25/AP50/recall numbers in Table 10 cannot be audited; if labels are YOLO-derived, the comparison against YOLOv8n would be systematically biased. Please provide the full annotation protocol, including inter-annotator agreement if applicable, and ideally release the labels or a public benchmark.","section":"Sec. 6.11, Table 10"},{"comment":"The field-sequence lock recall of 97.9% is measured only on YOLO-confirmed frames, which the paper states are close/mid-range frames. The distant regime, which covers 78% of the sequence and is exactly the regime the method targets, has no independent ground truth; Limitation 4 says so explicitly. As a result, the claimed advantage of the temporal representation over single-frame detectors is not directly validated in the operating regime where the paper argues it matters most. Please either hand-label a sample of distant frames, or clearly separate the claim into 'validated on close/mid range' and 'unvalidated on distant range'.","section":"Secs. 6.8, 6.12 and Limitation 4"}],"minor_comments":[{"comment":"The YOLOv8n row is formatted as '0.886 0.8710.103', which is ambiguous; it should be '0.886 / 0.871 / 0.103'.","section":"Table 5"},{"comment":"The classical diff+blob baseline is reported as AP25=0.120 in Table 4 but the precision-recall legend in Fig. 4 shows AP=0.119. Please reconcile the thresholds or state that the difference is due to the operating point.","section":"Fig. 4 / Table 4"},{"comment":"Several places render 'UAV' as 'UA V' (e.g., abstract and Section 6.11). Please fix the spacing.","section":"Abstract / body"},{"comment":"The verified false-lock rate of ~720/hr is based on only 2 minutes of screen-admitted footage and is dominated by one dense-vegetation video. The paper notes this, but a small confidence interval or per-video breakdown would help calibrate the KPI.","section":"Sec. 6.8"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically serious and internally careful, but the central transfer claim rests on synthetic pseudo-clips and an undocumented real-video labeling protocol. I would support publication after the authors either supply an independent real-video benchmark with hand labels, or substantially reframe the claims as conditional on the synthetic-motion model and add explicit sensitivity analysis. The reference list is clean and contains no self-citations; fit with a vision journal is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. It's a systems paper that actually delivers: a 22k-parameter detector over ego-motion-normalized residual channels runs at 31 FPS on a Pi Zero 2W and beats YOLOv8n by 2.7x AP25 on a 1000-video real-world corpus at 1/1100th the compute. The novelty is not any single component — LK+RANSAC ego-motion, compensated differencing, CenterNet heads are all established — but the assembly: analytically computed temporal evidence in the input, a deliberately stateless detector, and a learned verifier that tests whether the track moves like a human. They also document two int8 calibration failure modes that would otherwise reproduce as silently wrong models. That is useful for anyone deploying small quantized detectors.\n\nWhat makes me trust the paper more than the average CV submission: every ablation is over three seeds with std, the YOLOv8n baseline is fine-tuned on the identical training mixture, the quantization study is controlled, and the limitations section actually names the right next steps. The negative result on TSM is clean and acted on. This is serious engineering reporting.\n\nThe soft spot the stress-test flags is real. The held-out VisDrone evaluation converts static images into 4-frame pseudo-clips using the deployed Stage A/B code with hand-set trajectory parameters (up to 4 px/frame drift, 0.008 rad roll, 0.6% scale rate). The motion channel therefore encodes the authors' model of drone motion, not actual drone motion. In-domain tests share the same generative process, so they cannot validate the transfer. The one evaluation on genuine real videos (Table 10) reports AP but never states the labeling protocol; without that, the headline 0.462 AP25 is not fully auditable. That said, this is a limitation the authors openly concede (Limitations 2 and 4), and it does not undermine the core systems claim — deployment at 31 FPS on a milli-watt platform is measured directly. What it means is the generalization ratio (0.444 vs 0.103 on VisDrone) is believable but not proven.\n\nBottom line: this paper deserves a serious referee. I'd suggest the editor send it out with a specific request: get the Table 10 labeling protocol or drop the claim. For a reading group, it's a good case study in honest ablation design. I'd cite the quantization failure modes even if I skip the transfer claims.\n\nRecommendation: accept-shaped conditional — send to review, but the reviewer should push on the pseudo-motion assumption.","headline":"A genuinely clever systems paper with unusually honest ablations; the transfer claims rest on a synthetic-motion protocol that needs one more independent check.","tokens_in":23598,"tokens_out":2913,"would_cite":true,"duration_ms":26682,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Temporal motion evidence, not bigger networks, is what makes aerial person tracking work on milliwatt hardware.","keywords":["ego-motion compensation","temporal signatures","small-object detection","person tracking","int8 quantization","synthetic training curriculum","aerial video","edge deployment"],"falsifier":"Measure the deployed detector on a sizable corpus of real drone video with independently human-labeled person boxes, especially at 10-60 px target heights, without converting frames through the synthetic pseudo-clip procedure; if its AP25 there does not clearly beat an identically evaluated single-frame detector, the transfer claim is refuted.","tokens_in":1216,"feed_emoji":"🚁","tokens_out":4708,"duration_ms":71755,"temperature":0.7,"pith_summary":"The paper tries to show that on milliwatt-class drone hardware, person tracking does not need a bigger network; it needs the right input. By removing the camera's own motion with classical geometry and handing a tiny detector three input channels — luminance plus two ego-motion-normalized residual-motion maps — a 22k-parameter stateless network can detect people barely visible as 10-60 pixel smudges. The evidence is temporal, but it is computed analytically, not learned. The system generalizes to real drone imagery where a single-frame detector fine-tuned on the same data collapses, while running at real-time speed on a small single-board computer.","feed_headline":"Motion channels beat bigger networks on real drone footage","feed_subtitle":"A 7.6-MFLOP stateless detector tracks 10-60 px people at 31 FPS on milliwatt hardware and out-generalizes a 1,100x larger rival.","key_machinery":"The carrying mechanism is the ego-motion-normalized residual channel stack: each frame is warped by a 4-DoF similarity transform estimated from sparse optical-flow tracks, and the network receives luminance, residual-motion magnitude, and signed temporal difference. Under exact registration, static scenes vanish from the motion channels, so everything bright there is independently moving; registration noise becomes a structured nuisance the detector learns to discount. Because the same code generates these channels during training and deployment, the synthetic-to-real gap shrinks. Motion evidence is thereby supplied analytically for free, which is what lets a 7.6-MFLOP stateless detector rep","core_discovery":"Temporal evidence for detecting small moving people belongs in the input representation, computed by estimating and removing ego-motion, not in learned temporal machinery or appearance capacity. The deployed detector is stateless: adding temporal-shift modules lowers accuracy once the input channels carry ego-motion-normalized motion, and removing the motion channels barely changes in-domain score but collapses held-out real drone imagery 6x. Trained on synthetic clips whose motion channels were generated by the deployed ego-motion code, the int8 detector reaches 0.694 AP25 in-domain and 0.444 on held-out real drone stills, versus 0.103 for a far larger single-frame detector fine-tuned on th","pith_inferences":["If the transfer claim holds, the same recipe could apply to other small moving objects on moving platforms — animals or vehicles — provided their motion signatures can be synthesized in training.","The dependency on a synthetic-motion curriculum means the headline numbers are only as trustworthy as the hand-set trajectory parameters; an independently labeled real benchmark with distant-regime ground truth would be the decisive test.","The paper acknowledges that a stationary person starves the motion channels; this implies a hybrid mode that re-activates appearance evidence when motion is absent could extend coverage beyond the current design.","The verifier's 'moves like a human' test is trained on walk cycles; testing whether gait-phase articulation is enough to reject other periodic movers, such as animals, would sharpen the claim about what the motion signature actually captures."],"forward_implications":["Follow-me drones can track a person with a 22k-parameter detector plus classical geometry, at 31 FPS on milliwatt-class hardware, holding lock through occlusion.","Accuracy that transfers to real imagery comes from motion evidence, not appearance capacity: variants without motion channels collapse 6x on held-out real data, and fine-tuning a large single-frame detector on the same data does not save it.","Learned temporal modules are unnecessary and harmful once analytic motion channels are present; stateless inference is the right architecture.","int8 deployment is only free if calibrated correctly: moving-average range estimation silently halves accuracy, regression heads need true min-max ranges, and stateful graphs need propagated cache states.","A 1-D verifier of human-motion signatures can halve false positives on person-free footage, enabling a low false-alarm system."],"fun_headline_variants":["Put motion in input channels, not in network, to track tiny people from drones","Stateless 7.6-MFLOP detector beats 1,100x larger YOLOv8n on drone footage","Ego-motion normalization puts temporal evidence in input, enabling 31 FPS tracking","Temporal evidence is input, not learned: 7.6-MFLOP net tracks drone people at 31 FPS","Motion channels, not model size, make drone person tracking work at milliwatts"],"cache_read_input_tokens":24704,"weakest_assumption_plain":"The generalization result rests on the assumption that the hand-set synthetic-motion recipe — up to 4 px/frame drift, 0.008 rad/frame roll, 0.6% scale rate, and per-person lower-body shear applied to static imagery — reproduces the task-relevant motion statistics of real drone footage.","fun_headline_variants_meta":{"raw":{"variants":["Put motion in input channels, not in network, to track tiny people from drones","Stateless 7.6-MFLOP detector beats 1,100x larger YOLOv8n on drone footage","Ego-motion normalization puts temporal evidence in input, enabling 31 FPS tracking","Temporal evidence is input, not learned: 7.6-MFLOP net tracks drone people at 31 FPS","Motion channels, not model size, make drone person tracking work at milliwatts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001796,"raw_usage":{"total_tokens":7006,"prompt_tokens":929,"completion_tokens":6077,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":5952}},"tokens_in":673,"tokens_out":6077,"duration_ms":38901,"temperature":1.0,"reasoning_tokens":5952,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:44:14.560045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the deployed detector on a sizable corpus of real drone video with independently human-labeled person boxes, especially at 10-60 px target heights, without converting frames through the synthetic pseudo-clip procedure; if its AP25 there does not clearly beat an identically evaluated single-frame detector, the transfer claim is refuted.","supporting_citations":[],"review_version":1}